You are on page 1of 6

Insect Inspired Robots

John Lim1,2 , Chris McCarthy1,2 , David Shaw1,2 , Nick Barnes1,2 , Luke Cole2
1 2
Department of Information Engineering, RSISE, National ICT Australia,
Australian National University Australia

{johnlim, cdmcc, davids}@rsise.anu.edu.au, Nick.Barnes@nicta.com.au, cole@lc.homedns.org

Abstract Furthermore, given the limited processing power of the


insect brain, higher level behaviours such as perception
Insect behaviour has been a rich source of in- and scene understanding are also of interest - how does
spiration for the field of robotics as the per- a fly differentiate between a hand moving in quickly to
ception and navigation problems encountered swat it and a wall rushing in towards it as it attempts
by autonomous robots are faced also by in- to land on that wall?
sects. We work towards faster and more ro- It is clear that this area of bio-inspired visual naviga-
bust biologically-inspired, vision-based algo- tion is rich with potential research questions. The goal of
rithms for various tasks such as navigation and our research is to develop bio-inspired systems and algo-
control. In this paper, we introduce a research rithms that will solve fundamental navigation problems
platform, the Insectbot, that was built as a more efficiently and robustly. Presently, we are focusing
testbed for new algorithms and systems. Cur- on a subset of problems with a smaller scope than the
rent progress as well as the future direction of stated goal; we seek to concentrate our research on visual
our work is also presented. egomotion estimation, visual servoing, docking, feature
detection and corridor-centring. These are basic subsys-
1 Introduction tems that are integral within any autonomously moving
robot and the benefits of improved algorithms in these
Insects and other arthropods effortlessly navigate com- areas are many.
plex environments despite having relatively simple ner- The organization of this work is as follows. Section 2
vous systems. This is often attributed to robust and will introduce the robot platform that has been devel-
efficient motion control strategies where action and per- oped as a testbed for experimenting with new research
ception are closely coupled. Well-known visuo-motor be- ideas in a real-world environment. Then, Section 3 will
haviours include the use of optical flow for flight sta- detail several navigational subsystems developed on the
bilization, corridor-centring, flight odometry and the platform. These include flow-based corridor-centring be-
execution of smooth, grazing landings [Srinivasan and haviour and a flow-based docking strategy. Future work
Zhang, 2004], some of which have already been suc- and ongoing research will be discussed in Section 4 before
cessfully implemented in robots [Argyros et al., 2004; concluding.
Iida, 2003].
These motion strategies exploit the advantages of spe-
cialized eyes that provide sight over nearly the entire
2 Research Platform: InsectBot
viewsphere. It has been shown that a spherical field To facilitate the research interests and goals of insect in-
of view is advantageous to the task of estimating self- spired robotics, a novel mobile robot platform has been
motion [Fermuller and Aloimonos, 1998]. This may be a developed and appropriately named InsectBot. Two
contributing factor as to why flying creatures often have primary design features distinguish the InsectBot from
near-panoramic vision. many of the more common robotics research platforms
Insect navigation and control is also of interest to [ActivMedia, 2006; iRobot, 2006; Acroname, 2006].
robotics since it is believed that computationally cheap The first primary design feature is horizontal omni-
strategies are used as opposed to methods involving re- directional motion. The second primary design feature
construction and map-building. A classic example is bee- is vertical motion for a stereo camera system that uses a
inspired corridor-centring that balances the optical flow pair of hemispherical view (“fish-eye”) camera lens, each
between the left and right walls [Srinivasan et al., 1999]. providing a 190◦ field of view. These two design features
allow 4 degrees of motion, such as simulating the aerial (NICs); and more degrees of motion, such as a tilt and/or
landing of a honeybee. rotation mechanism for the vision system.
The horizontal omni-directional motion is performed Figure 1 shows a picture of the current InsectBot, how-
using four omni-directional wheels, spaced 90◦ apart. ever please refer to http://robot.anu.edu.au/∼insert for
A small number of mobile robots exist that use the full collection of the InsectBot photos, videos and
omni-directional wheels: The Palm Pilot Robot Kit research status.
[Carnegie-Mellon, 2001] was an early three wheeled
omni-directional robot and the Cornell Robocup [Cor-
nell, 2006] team used four wheeled omni-directional
robots, which may have contributed to their over-
whelming success at the international RoboCup champi-
onships. However the use of four omni-directional wheels
provides more stability for heavy payloads, which will
enable large payload capabilities for the InsectBot. Fur-
thermore the four omni-directional wheels provide an im-
proved and simplified motion control system [Ashmore
and Barnes, 2002].
The stereo cameras and their fish-eye lens have been
mounted so a small overlap exists between the cameras
in the visual field. This overlap coupled with almost 360◦
field of view provides a vision system that is quite com-
mon among insects within the natural world [Wehner,
1981]. This vision system is attached to a lift mechanism
allowing 300mm of vertical motion for the entire vision
system. Mounting this novel vision system to the omni-
directional base of InsectBot provides a mechanism for
simulating some key aspects of complex motions such
as flying. Helicopters and more recently the CSIRO
Cable-Array robot [Usher et al., 2004] are well known
approaches to studying flight. However helicopters are
naturally unstable and extremely hard to control. The
CSIRO Cable-Array robot can only operate within its
environment. Furthermore using cables to move the
cameras introduces wave-like effects which must be ac-
counted for in software. In contrast, the InsectBot can
operate in any environment with a flat surface, and has
acute control over its degrees of motion, allowing the
simulation of some key aspects of flight.
The InsectBot is designed to fit any ATX style moth-
erboard and supports most sensors, such as a SICK laser
range finder. Currently a Mini-ITX VIA EPIA SP1300
motherboard with 1GB of DDR400 RAM is used for
communication to the motors and sensors along with Figure 1: The InsectBot
handling the motion control of the drive system and
lift mechanism. Visual processing tasks are handled off
board through one or more high end computers. Com-
3 Navigational Subsystems
munication between these high end computers and the 3.1 Flow-based Corridor Centring
InsectBot is performed through a 56kbps wireless con- Corridor centring, inspired by observations in honeybees
nection or a 100Mbps wired connection (if required). [Srinivasan et al., 1999], can be achieved by adjusting the
Approximately, 30 Amp-hours at 24V is available, which robot’s heading so as to maintain a balance of flow mag-
with the current InsectBot gives a few hours run time. nitude in the outer regions of the robot’s view. We have
Future technical developments on this robot will include: implemented a corridor-centring subsystem that makes
increased communication bandwidth through the use of use of full optical flow estimation over images from a sin-
muliple wireless and/or wired Network Interface Cards gle forward facing camera to achieve this. We emphasise
Multiple centring trials were conducted using a
straight 2.5 metre long corridor with heavily textured
walls (see Figure 2). An overhead camera was used to
track the robot’s path, and from this, an assessment of
directional control stability was conducted.
Results indicated that in addition to the choice of op-
tical flow method, the choice of spatio-temporal filter
(for gradient-based methods) also had a significant ef-
fect on the overall performance of the system. Most
notably, the temporal delay introduced through frame
buffering was particularly problematic. The best over-
all performance was achieved using Lucas and Kanade’s
gradient-based flow method [Lucas and Kanade, 1984] in
conjunction with Simoncelli’s multi-dimensional paired
filters [Simoncelli, 1994].
Figure 2: Corridor centring workspace 3.2 Flow-based Docking
For a mobile robot to interact with an object in its envi-
the need to consider such navigation subsystem’s in the ronment, it must be capable of docking in close proximity
context of the broader system they inhabit, and the di- with the object’s surface. Of particular importance is the
verse range of tasks the system needs to perform. For control of the robot’s deceleration to an eventual halt,
this reason, we use full optical flow estimation in the im- close enough to the object for the interaction to take
plementation of all flow-based subsystems, rather than place. To achieve this, the robot must acquire a robust
cheap approximations such as normal flow, or planar estimation of time-to-contact, and from this, control the
models. In so doing, this raises the important question velocity accordingly.
of which optical flow method to choose for flow-based For a single camera approaching an upright surface,
navigation subsystems such as corridor centring. a common method of estimating time-to-contact is to
While previous optical flow comparisons assess the ac- measure the image expansion induced by the apparent
curacy and efficiency of optical flow techniques, they do motion of the surface towards the camera. This can be
not adequately support a systematic choice of technique obtained from the optical flow field divergence. The use
when considering which to choose for a real-time naviga- of visual motion to gauge time-to-contact is well sup-
tion system. To address this shortcoming, we have con- ported by observations in biological vision. [Srinivasan
ducted a comparison of optical flow techniques specifi- et al., 2000] observes how honeybees use visual motion to
cally for flow-based navigation. A preliminary study con- decelerate and perform smooth graze landings. Lee [Lee,
sidered the choice of spatio-temporal filter for use with 1976] theorised that a human driver may visually con-
gradient-based optical flow estimation for tasks such as trol vehicle braking based on time-to-contact estimation
corridor centring and visual odometry [McCarthy and obtained from image expansion.
Barnes, 2003]. Subsequent work has extended this to an In much of the previous work with divergence-based
examination of several well cited optical flow methods time-to-contact estimation, divergence is measured at
as well [McCarthy and Barnes, 2004]. Unlike previous the same image location each frame. Such strategies ig-
comparisons, we emphasise the need for in-system com- nore the effect of focus of expansion (FOE) shifts on the
parisons of optical flow techniques, in the context of the divergence measure across the image. Robot egomotion
tasks they need to perform. is rarely precise, and even where only translation is in-
As part of this comparison, each optical flow method tended, rotations will be present due to steering control
was integrated into the control loop of the corridor cen- adjustments, differing motor outputs, bumps, and noisy
tring subsystem of a mobile robot. Directional control flow estimates.
for corridor centring was achieved using the simple con- We have developed a robust strategy for docking a
trol law: mobile robot in close proximity with an upright surface,
θ = K(τl − τr ), (1) that accounts for small rotational effects through the
constant tracking of the FOE [McCarthy and Barnes,
where τl and τr are the average flow magnitudes in the 2006]. Through theoretical derivation and experimen-
left and right peripheral views respectively, and K is tal results, we show that computing the time-to-contact
a tuned proportional gain. θ is then used directly for with respect to the location of the FOE rather than the
directional control. image centre, accounts for shifts of the optical axis due
inspired systems that are being investigated.
The detector locates the likely 2D projection of rect-
angles in the environment by finding quadrilaterals with
sides aligning with detected vanishing lines within the
image in an order of complexity suitable for real-time
applications (refer to Figure 5). This detector has been
designed primarily for robot visual mapping.
A brief outline of the stages of the perspective rectan-
gle detection algorithm is given below [Shaw and Barnes,
2006]:
1. Vanishing Point Detection: Estimate the position
and type of vanishing points and lines within the
image.
2. Perspective Oriented Edge Detection: Determine
Figure 3: Setup for on-board docking tests. the directional components of the gradient aligned
with the vanishing points.
On-board Docking - L&K Recursive 3. Line Segment Detection: Estimate the line seg-
0.45 ments that are potential quadrilateral sides based
on the perspective oriented edges.
0.4
robot forward velocity

4. Quadrilateral Detection: Determine the quadrilat-


0.35
erals from the intersection of detected line segments.
0.3
A benefit of using rectangular features is that they rep-
0.25 resent implicit structural information (such as the size of
0.2 Trial 1 wall panels and doors) that is of benefit for robotic appli-
Trial 2 cations. Even with a simple matching metric, significant
0.15 Trial 3
Trial 4 structural information about a scene can be derived.
90 80 70 60 50 40 30 20 10 0
distance to wall (cm) 4 Future Work
In this section, we identify key areas of interest for future
Figure 4: On-board velocity-distance profiles for docking
investigation.
trials.
The common motion control algorithms in robots are
often computationally expensive because they involve re-
to instantaneous rotations. construction, map-building, geometric models, estima-
To test the FOE-based docking strategy, it was im- tion of scene depth or a computation of the homography
plemented on-board a mobile robot with a single for- between views. These methods are also less robust to
ward facing camera. Over multiple trials, it approached errors in camera and robot calibration.
a heavily textured, roughly fronto-parallel wall, attempt- In contrast, the limited processing power of the neural
ing to safely stop as close to the wall as possible without structure of insects suggests that insects abandon such
collision (see Figure 3). methods in favor of simpler and more robust strategies to
Figure 4 shows velocity-distance profiles for four of solve the same problems. Motion control in bees [Srini-
the trials conducted. In all four trials, the FOE-based vasan et al., 1999] and the use of landmarks for visual
strategy was used by the robot to dock in close proximity piloting in desert ants [Lambrinos et al., 2000] are oft-
to the surface without collision. cited examples. Also, it is shown in [Bekris et al., 2004]
that any position on the ground plane can be reached
3.3 Perspective Rectangle Detection using a navigational algorithm based purely on the an-
Robust feature detection is an important subsystem for gles between image features. We believe that there re-
many navigational tasks. In many indoor scenes, the mains potential for many other computationally cheap
environment consists predominantly of rectilinear struc- yet powerful control algorithms to be derived from bio-
tures. To capitalise on this for navigation, we have de- logical solutions.
veloped a detector for finding these perspective rectangle Furthermore, we intend to integrate these various al-
structural features that runs in real-time. This detector gorithms into a single, unified navigational system with
is expected to be an important component for the bio- the capabilities for a large range of navigational tasks.
Figure 5: Perspective rectangle features in a corridor sequences.

This will enable us to study the behaviour of the system be able to identify different scenarios and environments
as one or more motion strategies work in concert. This and engage the corresponding subsystems in order to
may lead to new synergies between existing algorithms complete the necessary tasks.
being discovered.
References
5 Conclusion [Acroname, 2006] Acroname, 2006.
There are often many solutions to the same fundamental http://www.acroname.com.
problems of robot vision and motion. Traditional solu- [ActivMedia, 2006] ActivMedia, 2006.
tions to these work well in general but become difficult http://www.activrobots.com/ROBOTS.
to implement when the requirement is real-time perfor- [Argyros et al., 2004] A.A. Argyros, D.P. Tsakiris, and
mance. Since autonomous robots and living creatures
C. Groyer. Bio-mimetic centering behavior: Mobile
encounter parallel problems as they tackle similar tasks,
robots with panoramic sensors. IEEE Robotics and
researchers have sought to derive new insights from the
Automation Magazine, Special Issue on Panoramic
biological solutions to those problems in order to develop
Robotics, 11(2):21–30, 2004.
new solutions that work robustly in real-time.
We have presented several visual navigation strategies [Ashmore and Barnes, 2002] M. Ashmore and
which were implemented on the research platform. As N. Barnes. Omni-drive robot motion on curved
highlighted in Section 4, various other research areas will paths: The fastest path between two points is not
also be investigated. All of these represent subsystems a straight-line. Australian Joint Conference on
within an autonomous robot. As they come together, Artificial Intelligence, pages 225–236, December 2002.
research will then concentrate on the development of [Bekris et al., 2004] K. E. Bekris, A.A. Argyros, and
higher-level perception and cognition systems that will L. E. Kavraki. Angle-based methods for mobile
robot navigation: Reaching the entire plane. In In- Image Proc, volume I, pages 790–3, Austin, TX USA,
ternational Conference on Robotics and Automation 1994. IEEE Sig Proc Society.
(ICRA04), pages 2373–2378, 2004. [Srinivasan and Zhang, 2004] M. V. Srinivasan and
[Carnegie-Mellon, 2001] Robotics Institute Of Carnegie- S. W. Zhang. Visual motor computations in insects.
Mellon. Palm pilot robot kit (pprk), 2001. Annual Review of Neuroscience, 27:679–696, 2004.
http://www.cs.cmu.edu/ pprk/. [Srinivasan et al., 1999] M.V. Srinivasan, J.S. Chahl,
[Cornell, 2006] Robocup Team At Cornell, 2006. K. Weber, S. Venkatesh, M.G. Nagle, and S.W. Zhang.
http://robocup.mae.cornell.edu. Robot navigation inspired by principles of insect vi-
[Fermuller and Aloimonos, 1998] C. Fermuller and sion. Robotic Autonomous Systems, 26(2-3):203–216,
Y. Aloimonos. Ambiguity in structure from mo- 1999.
tion: Sphere versus plane. International Journal of [Srinivasan et al., 2000] M. V. Srinivasan, S. W. Zhang,
Computer Vision, 28:137–154, 1998. J. S. Chahl, E. Barthand, and S. Venkatesh. How
[Iida, 2003] F. Iida. Biologically inspired visual odome- honeybees make grazing landings on flat surfaces. Bi-
ter for navigation of a flying robot. Robotics and Au- ological Cybernetics, 83:171–183, 2000.
tonomous Systems, 44(3-4):201–208, 2003. [Usher et al., 2004] Kane Usher, Graeme Winstanley,
[iRobot, 2006] iRobot, 2006. http://irobot.com. Peter Corke, Dirk Stauffacher, and Ryan Carnie. A
cable-array robot for air vehicle simulation. Aus-
[Lambrinos et al., 2000] D. Lambrinos, R. Moller, tralasian Conference on Robotics and Automation
T. Labhart, R. Pfeifer, and R. Wehner. A mobile (ACRA), December 2004.
robot employing insect strategies for navigation.
[Wehner, 1981] R. Wehner. Handbook of Sensory Phys-
IEEE Transactions on Robotics and Autonomous
Systems, 30(1):39–64, 2000. iology. Berlin/Heidelberg: Springer-Verlag, 1981.
[Lee, 1976] D. N. Lee. A theory of visual control of brak-
ing based on information about time to collision. Per-
ception, 5(4):437–59, 1976.
[Lucas and Kanade, 1984] B. Lucas and T. Kanade. An
iterative image registration technique with an applica-
tion to stereo vision. In Proceedings of DARPA Image
Understanding Workshop, pages 121–30, 1984.
[McCarthy and Barnes, 2003] C. McCarthy and
N. Barnes. Performance of temporal filters for
optical flow estimation in mobile robot corridor
centring and visual odometry. In Proceedings of
the 2003 Australasian Conference on Robotics and
Automation, 2003.
[McCarthy and Barnes, 2004] C. McCarthy and
N. Barnes. Performance of optical flow tech-
niques for indoor navigation with a mobile robot. In
International Conference on Robotics and Automation
(ICRA Q04), pages 5093–5098, 2004.
[McCarthy and Barnes, 2006] C. McCarthy and
N. Barnes. A robust docking strategy for a mo-
bile robot using flow field divergence. In IEEE/RSJ
International Conference on Intelligent Robots and
Systems (IROS 2006), page in press, 2006.
[Shaw and Barnes, 2006] David Shaw and Nick Barnes.
Perspective rectangle detection. In Proceedings of the
Workshop of the Application of Computer Vision, in
conjunction with ECCV 2006, pages 119–127, 2006.
[Simoncelli, 1994] E P Simoncelli. Design of multi-
dimensional derivative filters. In First Int’l Conf on